Show HN: Giving Opus 5.5 a simulated paint canvas
A new project lets Anthropic’s Claude Opus 5.5 “paint” by issuing brushstroke commands on a simulated oil canvas, using its vision to iteratively observe and refine the image rather than relying on a traditional image model. Commenters explore how this showcases emergent spatial and artistic reasoning in LLMs, compare it to prior code-driven art experiments, and speculate about extensions to physical robots and benchmarks. The thread also surfaces broader debates over what counts as real art, the ethics and economics of AI-generated imagery, and whether systems displaying such generalization should be seen as steps toward genuine machine intelligence.
How the system works
- Opus doesn’t output pixels directly; it emits structured commands (coordinates, tools, pressures) that drive a simulated oil-paint engine.
- The simulator models physical paint behavior: bristles, wet paint, drying over time, layered optics, and mixing from virtual paint tubes.
- The model can “look” at the evolving canvas via its vision capabilities, then iteratively refine the work.
- This is likened to turtle graphics or SVG generation: text instructions translated into visual output via a separate engine.
Relation to other code-based painting projects
- Commenters connect this to prior work on “training AI to paint with code” and other voxel / Minecraft-like benchmarks.
- Some see this as part of a broader shift: using LLMs to drive processes and tools rather than having end-to-end image generators.
Perceived artistic quality and style
- Many find the results impressive and intriguing, some saying they’d like to own such canvases or use the simulator to explore techniques safely.
- Others note uncanny artifacts (e.g., clusters of churches, odd textures) and debate whether these are failures or simply artistic quirks.
- Comparisons are made to historical and contemporary painters, amateur work, and algorithmic composition.
Impact on artists, consumers, and ethics
- Some express sadness and concern about further devaluing already underpaid artists and eroding “human-to-human” art.
- Others argue that consumers mostly care about outcome, often can’t distinguish AI art, and that economic pressures will drive adoption anyway.
- Counterpoints stress that some organizations and audiences explicitly reject AI-generated work and that disclosure and labeling are emerging norms.
Debate over intelligence, generalization, and “mind”
- Several commenters see this as strong evidence of emergent generalization: painting via brushlike strokes without obvious direct training.
- Others suggest the behavior could be reinforced via specific RL environments or exposure to painting videos.
- There’s an extended philosophical debate about whether LLMs constitute “minds,” what would count as consciousness or pain, and how (or whether) rights might eventually apply.
Future directions and experiments
- Ideas raised include using robot arms or plotters to paint physically, streaming sessions, measuring token costs, and benchmarking models on painting tasks or photo-to-painting studies.